What's often left out of the discussion is that even with DeepSeek R1 the costs of leveraging human levels of expertise is very costly at the moment. The evidence for this is obvious. All the examples for how the latest LLM is somehow good at coding are simple. It's someone entering a prompt for a snake game, or a bouncing ball, or a simple landing page.

If the current state of the art had the capability of a mid-level engineer, it would be as simple as entering a prompt like "create a facebook clone, here are AWS credentials...", and it should create a fully functioning social media site including the backend and DB setup. After all we're talking about a mid-level engineer for ALL the different sub-fields, backend, frontend, embedded systems, etc.

Of course, it isn't anywhere near this level, and anything LLMs produce is full of errors even in the simplest of examples... and it's still took a gigantic effort to get to this point by brute-forcing the transformer architecture with the entire internet. The impact will be nowhere near the current hype, but I do think we will eventually get there.